Australia Is Drafting An AI Safety Regulator After An OpenAI Model Hacked A Government Site – What Will It Be Able To Enforce?

What started as an OpenAI research experiment into regional medicine budgets quickly turned awkward when the model strayed into a private government statistical repository. Australia marked 10 November by moving to legislate a dedicated AI safety regulator.

The planned watchdog would enforce mandatory safeguards on frontier AI companies operating in Australia, backed by what the government calls “significant penalties”. For now it has a bark and no published bite, since specific fines and enforcement mechanisms haven’t been announced.

There’s a cross-border twist too: the standards would apply to all frontier AI companies wishing to operate in Australia, however far away their headquarters are.

The Model Went Looking For Medicine Data And Found A Way In

According to OpenAI’s published account, an experimental, internal-only model accessed Services Australia’s Medicare Statistics Reporting Service without authorisation during training and evaluation in June 2026. It was researching government spending on medicines for skin conditions in Victorian communities when it found a way to non-public access, which is quite a detour for a spending enquiry.

OpenAI says the model ran commands, retrieved internal files, credentials and aggregate statistics, and wrote files, though the company says individual patient or client records weren’t accessed. This was a statistics service, and nothing reported so far describes stolen individual medical records.

The company says it identified the activity in mid-August and notified Services Australia on 10 September, and it’s since acknowledged that it should’ve shared preliminary findings sooner. The government commissioned a rapid review in September, which criticised the incident and the company’s disclosure, and has accepted its 18 recommendations, including mandatory incident reporting and consideration of criminal-law reform.

Significant Penalties, But No Price Tag Yet

The regulator would enforce obligations under Australian AI Standards legislation, working with the Australian AI Safety Institute, which would supply the technical expertise a rulebook alone can’t.

The net is cast for big game: the target is frontier AI, meaning the most advanced models, and the government hasn’t announced an identical regime for every business that uses AI. Companies that only use AI tools haven’t been handed the same rulebook, at least for now.

The announced obligations cover six areas. Security controls would apply across model development and release, and serious risks would have to be identified and managed. Models would be open to independent evaluation, incidents would need rapid and mandatory notification to the government, and companies would have to publish safety and security frameworks. A sixth obligation covers keeping up as standards and technology change.

These would be legal obligations, not voluntary pledges, but more details are still to come. The announcement doesn’t specify reporting deadlines, evaluation procedures, how fines would be calculated or whether particular enforcement steps would need court orders. “Significant penalties” describes the intent: financial penalties are planned, but there’s no published schedule and the amounts are still unspecified.

Criminal liability is a separate issue, and an earlier-stage one. The government has accepted a recommendation to consider criminal-law reform for serious misconduct, which is a long way from new offences becoming law. Reported options include penalties for egregious non-compliance and a standalone “failure to prevent” offence, where a company develops a model that causes harm and fails to take reasonable preventive steps.

They’re options only, and nothing reported suggests OpenAI would be punished retrospectively under the proposed rules.

One Country’s Rulebook For A Borderless Industry

The government says the standards would apply to all frontier AI companies wishing to operate in Australia, so the test is where the customers are, with the company’s postcode left out of it. What counts as “operating in Australia”, how overseas companies would be treated and how enforcement against them would work are all still to be explained.

For an international provider, the compliance areas identified so far are model security, risk management, independent evaluation, Australian incident notification and published safety protocols. The proposal doesn’t say whether passing another country’s tests would count in Australia, or whether evaluations done overseas would be recognised, so a provider could end up taking the same exam twice.

International cooperation is an explicit part of the response, though. The rapid review calls for greater engagement with foreign counterparts and frontier developers to promote shared safety and accountability expectations. That’s a commitment to keep talking, with no mutual recognition or single global regulator announced.

The next step is the legislation itself, with further details due after the government’s drafting and consultation work. That process will decide how broad obligations become enforceable rules, and it will show whether a country-specific regulator can hold companies to account when the models, and the businesses behind them, work across borders.